Training Models on Office Work Improves Coding by 5.8%
echen · x · 2026-08-05
Surge AI discovered that post-training a model solely on office tasks (spreadsheets, docs) unexpectedly improved its software engineering skills, boosting SWE-Bench Pro performance by 5.8 percentage points.
The researchers likened this to a junior engineer leveling up after planning weddings. They found the model developed a Goal-Directed Execution capability:
- Forming the right local goals
- Building an accurate picture of the environment
- Staying faithful to higher-level objectives
- Verifying goal achievement
These underlying skills successfully transferred across different agent workflows.
Related event: Surge AI Finds Office Task Training Boosts Coding Skills(2 posts)→
More from coding & agent
- Bret Victor's Classic Essay: Rethinking Programming and State — mmmbchang · 2026-08-05
- Logfire vs Braintrust: Agent Eval Pricing Shows a 700x Cost Difference — samuelcolvin · 2026-08-05
- Dev Explains Why Coding Agents Fail: Cannot Observe or Manipulate State — mmmbchang · 2026-08-05
- Developer Shares Journey of Building a 100% Private 'Second Brain' AI Agent — FlolightC · 2026-08-05
- Perfect Data QA is the Key Unlock for Recursive Self-Improvement — simonguozirui · 2026-08-05
- Voice AI Agent Evals and the Consumer Market Opportunity — agihouse_org · 2026-08-05